Semantic Segmentation for Automated Identification of Bacterial Colonies

The images used in this work depict bacterial colonies grown on agar in Petri dishes. In total, 8 bacterial strains were imaged. The dataset consists of 111 images. The dataset was acquired using a custom-built linescanning hyperspectral imaging setup combining an Im-Spector V10e imaging spectrograph and a 5 MP CMOS camera, with uniform broadband and near-infrared LED illumination covering 400-1000 nm. The field of view was acquired by motorized scanning. In this case, reflectance was captured with a spectral resolution of .3 nm under white light illumination. Each hyperspectral image has a spatial resolution of 1800 × 1224 pixels (width × height) and a spectral dimension of 2047 bands, meaning that each spatial pixel contains intensity values across all 2047 captured wavelengths. Such hyperspectral datasets are commonly referred to as hyperspectral image cubes. Imaging was performed in multiple sessions distributed over several weeks across multiple months with different age of colonies. The hyperspectral images were normalised using a white reference and converted into three-channel RGB images for visualization and dimensionality reduction. After the hyperspectral to RGB conversion, semantic masks were drawn manually on the RGB images. Semantic annotations were later verified by a microbiologist. The data was split into train, validation and test subsets with stratification, where each bacterial strain is represented in each subset. Train subset contains 85 images, validation subset contains 15 images and test subset contains 11 images. Please cite the conference paper... (will later be disclosed)